Collaborative Research: III: Small: A DREAM Proactive Conversational System

会话系统，如智能助手，在我们的许多日常设备和环境中都可以找到，包括电话、智能扬声器和各种在线服务。然而，这些系统大多是反应性的。它们只响应用户明确的请求。在这样做的过程中，他们依赖于用户提出正确的问题，这对那些信息素养低的人来说是不利的。此外，涉及探索性搜索、头脑风暴和复杂决策的情况需要双方积极参与的对话。该项目将开发一种智能会话代理，它不仅可以被动地回答问题，还可以进行双向、主动的对话，与用户进行头脑风暴，从正在进行的对话中产生新的见解，并为共享对话做出新的贡献。这种智能会话代理将建立在基于离线分层模型的深度强化学习框架上。该算法框架的分层设计将提供高层次的会话策略和实际的话语生成。它将允许智能体动态检测并决定何时插话，将对话带入一个新的路径，丰富知识和主题，管理多层任务目标，并评估潜在的机会和陷阱，以提出主动响应。这种新颖的方法还将包括状态跟踪和建模，这将结合对会话环境模型的理解，主动行动计划，以扩展当前会话任务的知识和主题，并提出潜在的行动，以及批评和验证提议的响应，这将有助于识别机会和陷阱，从而增强主动智能。系统的评估将涉及项目不同阶段的客观和主观度量。在验证阶段，将使用现有数据集进行离线实验，以评估和微调新的学习框架。接下来将是众包阶段，用于生成基础真相并评估对话的新颖性和适当性。最后，用户研究阶段将评估系统的有效性，以及用户体验和任务性能。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Conversational systems, such as smart assistants, are found in many of our everyday devices and contexts, including phones, smart speakers, and various online services. However, most of these systems are reactive. They only respond to what a user asks explicitly. In doing so, they rely on the user asking the right question, which disadvantages those with low information literacy. In addition, situations involving exploratory search, brainstorming, and complex decision-making require a conversation with both sides actively engaged. This project will develop an intelligent conversational agent that can not only reactively answer questions but also conduct two-way, proactive conversations, brainstorm with the user, generate new insights from the ongoing conversation, and make novel contributions to the shared dialogue.Such an intelligent conversational agent will be built on an offline hierarchical model-based deep reinforcement learning framework. This algorithmic framework's hierarchical design will provide high-level conversation strategies and the actual utterance generation. It will allow the agent to dynamically detect and determine when to chime in to take the conversation to a novel path, enriching knowledge and topics, managing multi-tiered task objectives, and assessing potential opportunities and pitfalls to propose proactive responses. This novel approach will also include state tracking and modeling that will incorporate an understanding of the conversational environmental model, proactive action planning to expand the current conversational task's knowledge and topics and proposes potential actions, as well as critiquing and validating the proposed responses that will help recognize opportunities and pitfalls to empower proactive intelligence. Evaluation of the system will involve both objective and subjective measures at the project's different phases. During the validation phase, offline experiments with existing datasets will be performed to assess and fine-tune the new learning framework. This will be followed by the crowdsourcing phase for generating ground truth and evaluating novelty and appropriateness of the conversations. Finally, the user studies phase will evaluate the system's effectiveness, as well as user experience and task performance.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.